A web application for training and generating images with GANs (generative adversarial networks), which was a popular method for image-based generative AI from its development in 2014, until diffusion kind of took over around 2020. Given DALL-E and Stable Diffusion, this technique is now mostly obsolete but it was fun while it lasted.
This was my major project for software design and development in high school in 2018. So the disclaimer is that I probably don't agree with many of the decisions anymore (especially the one where it was built from scratch in vanilla JavaScript).
At the time this was probably the largest programming project I had completed.
All I had was a hammer, so the entire project was written from scratch using vanilla JavaScript. No deep learning libraries, no matrix multiplication, just rawdogging JavaScript with as its creator intended. I was quite literally doing a for loop over every neuron and summing up activations. Even the UI was made with rectangle/text primitives and powered by click events that I would listen for on a HTML Canvas.
Feynman's test for understanding a concept was to try to explain it to a 5 year old. I'd like to propose an alternative where you must build it from scratch in vanilla JavaScript.
The source code can be found on Github.
Application embedded below if you want to try it yourself, although I won't make any guarantees about the experience.
You can draw the data manually (not recommended). Otherwise the MNIST training data for each digit can be found on GitHub.